Karol Herbst at Red Hat who has been a longtime open-source NVIDIA driver contributor as well as Rusticl developer presented at XDC2025 on the NVK Vulkan driver’s cooperatrive matrix extension support. It turns out this Vulkan extension for machine learning / AI is proving fairly competitive with the open-source NVK driver compared to NVIDIA’s official driver stack.
VK_KHR_cooperative_matrix was introduced back in mid-2023 as part of the efforts for enhancing the Vulkan API around AI/ML workloads. With time there’s been a lot of work on making the Vulkan API suitable for AI workloads and even NVIDIA finding success and competitive performance compared to CUDA.
During a lightning talk at XDC2025 in Vienna, Karol Herbst presented on NVK’s support for cooperative matrix. The one takeaway for end-users is the performance with NVK.
Karol found that the Vulkan cooperative matrix support on NVK started out at around 20% the speed of the official NVIDIA driver. But it’s reached 70% on the main Mesa development branch. Or most excitingly is with the latest development code, upwards of around 92% the speed of the official NVIDIA driver.
Those wanting to learn more can view Karol’s presentation above as well as the PDF slide deck.
